Professional Documents
Culture Documents
, (1)
where A and are real and constant ( being smaller
than zero) and x is a variable. For instance g(x)
could represent the distribution D(s) of the size s of
events in an experiment, or the power spectrum
P( f ) of a signal as a function of its frequency f. This
type of function is sometimes refered to as a power
law because of the exponent . By taking the log of
both sides, one obtains :
log g(x) llog Ajlog x. (2)
When plotted on a loglog scale, this type of function
therefore gives a characteristic straight line of slope
, which intersects the ordinate axis at log A (see Fig.
1 for example). When trying to t a power law to
experimental data, as I will do often in this review,
it is customary to rst take the log of the measure-
ments, and then to t a straight line to it (by the
least-square method for instance). This method
proves less susceptible to sampling errors.
Power laws are interesting because they are scale-
invariant. We can demonstrate this fact by changing
x for a new variable xh dened by x la xh, where a is
some numerical constant. Then replacing in equa-
tion (2), one gets :
g(axh) lA(a xh)
l(Aa
) xh
. (3)
The general form of the function is then the same as
before, i.e. a power law with exponent . Only
the constant of proportionality has changed from A
to Aa
(4)
where ` means is proportional to.
(2) Fractals in space
As we briey mentioned in the introduction, the
earths crust is full of structures without any
characteristic scale. This is, in fact, a property shared
by many objects found in nature. It was, however,
only in the 1960s and 1970s, with the pioneering
work of Mandelbrot (1977, 1983), that this fact was
given the recognition it deserved. We refer the
reader to Mandelbrots beautiful and, sometimes,
challenging books for an introduction to this new
way of seeing and thinking about the world around
us. Here, we will barely scratch the surface of this
very vast subject by focusing on the concept of
fractal dimension which will be useful to us later on.
Early in one of his books (Mandelbrot, 1983),
Mandelbrot asks the following, and now famous,
question: How long is the coast of Britain? The
intuitive way to answer that query is to take a map
of Britain, a yardstick of a given length d, and to see
how many times n(d) it can be tted around the
perimeter. The estimation of the length L(d) is then
din(d). If we repeat this procedure with a smaller
yardstick, we expect the length to increase a little
and nally, when the yardstick is small enough, to
ultimately converge towards a xed value: the true
length of the coast. This yardstick method, which is
in fact just ordinary triangulation, works well with
regular or euclidian shapes such as a polygon or a
circle. However, as Mandelbrot (1983) noticed,
triangulation does not bring the expected results
when applied to computing the length of coasts and
land frontiers. As we reduce the size d of the
yardstick, more details of the seashore or frontier
must be taken into account, making n(d) increase
quickly. It does so fast enough that the length L(d)
ldin(d) keeps increasing as d diminishes. Fur-
A B C D
(1)
(3)
(4)
(5)
10
3
10
2
10
1
Yardstick length d (km)
10
4
10
3
M
e
a
s
u
r
e
d
l
e
n
g
t
h
L
(
d
)
(
k
m
)
(2)
(6)
E
Fig. 2. (AD) Computation of the length of a circle of
radius 1 using the box counting method. A square lattice
of a given size d is laid over the curve. The number of
squares N(d) necessary to cover the circles perimeter is
then counted and the length of the circle L(d) approxi-
mated as diN(d). (A) d l0.5, N(d) l16 and L(d) l8.
(B) d l0.25, N(d) l28 and L(d) l7. (C) d l0.125,
N(d) l52 and L(d) l6.5. (D) d l0.0625, N(d) l100
and L(d) l6.25. The true value of the perimeter of the
circle is of course Ll26.28319. (E) Estimation using
the yardstick method of the length L(d) of the coast of
Australia (1), South Africa (3) and the west coast of
Britain (5), as well as the land frontiers of Germany (4)
and Portugal (6) as a function of the yardstick d used to
make the evaluation. The length of a circle (2) of radius
1500 km is also included for comparison. Reproduced
from Mandelbrot (1983).
thermore, in order to get a better estimate, one
should use a map of increasing resolution, where
previously absent bays and subbays, peninsulas and
subpenisulas now appear. Taking into account these
new features will also increase the length L(d) even
more.
Instead of using triangularization, one can rely on
a similar and roughly equivalent method called box
counting (Mandelbrot, 1977, 1983). It is more
practical and signicantly simpler to implement on a
computer. One superimposes a square grid of size d
on the curve under investigation, and counts the
minimum number N(d) of squares necessary to cover
it. The length of the curve L(d) is then approximated
by diN(d). Fig. 2AD illustrates the procedure for
a circle of radius 1. The estimate is at rst a little o,
mostly because of the curvature of the circle.
However, as d gets smaller this is quickly taken into
167 Scale invariance in biology
account and the measured length converges toward
the true value Ll26.28319. Applying these
methods to other curves like the coast of Britain gives
dierent results. Fig. 2E shows the variation of L(d)
as a function of d for dierent coastways and land
frontiers. As can be seen, the data for each curve
followa straight line over several orders of magnitude
of d. This suggests the power-law parametrization of
L(d) (Mandelbrot, 1977, 1983):
L(d) `d"
(5)
where is some real parameter to be tted on the
data.
As expected the perimeter of the circle quickly
converges to a value, and stays there for any smaller
values of d. For this part of the curve L(d), l1 (a
horizontal line) ts the data well. The same goes in
Fig. 2E for the coast of South Africa (line 3).
However, for all the other curves, L(d) follows a
straight line with non-zero slope. For instance, in the
case of the west coast of Britain, the line has slope
-k0.25, and therefore L(d) `d
!.#& and 1.25.
Mandelbrot (1977, 1983) dened as the number
of dimensions (or box dimension when using the box
counting method) of the curve. For the circle, l
1 as L(d) is independent of d: we recover the intuitive
facts that a circle is a curve of dimension 1, with a
nite value of its perimeter. The same is also almost
true for the coast of South Africa.
However, for the coast of Britain for instance, is
not an integer, which indicates that the curve under
investigation is not Euclidian. Mandelbrot coined
the term fractal to designate objects with fractional
or a non-integer number of dimensions. Also, the
data in Fig. 2E indicate that Britain possesses a coast
with a huge length, which is best described as quasi-
innite: L(d) `d
!.#& goes to innity as d goes to
zero. This is due to the fact that no matter how
closely we look at it, the coastway possesses structures
such as bays, peninsulas, subbays and subpeninsulas,
which constantly add to its length. It also means that
no matter what scale we use, we keep seeing roughly
the same thing: bays and peninsulas featuring
subbays and subpeninsulas, and so on. The coastway
is therefore eectively scale-invariant. Of course, this
scale invariance is not without bounds : there are no
features on the coastway larger than Britain itself,
and no subbay smaller than an atom. We also note
that the more intricate a curve, the higher the value
of its box dimension : a curve which moves about
so much as to completely ll an area of the plane will
have a box dimension 2.
The concept of non-integer dimension might seem
a little strange or articial at rst, but the geo-
metrical content of the exponent is not. It is a
measure of the plane- (space-)lling properties of a
curve (structure). It quanties the fact that fractals
with a nite surface may have a perimeter of (quasi)
innite length. Similarly, it shows how a body with
nite volume may have an innite area. It is
therefore not surprising to nd fractal geometry in
the shape of cell membranes (Paumgartner, Losa &
Weibel, 1981), the lungs (McNamee, 1991) and the
cardiovascular system (Goldberger & West, 1987;
Goldberger, Rigney & West, 1990). Fractal ge-
ometry also helps explain observed allometric scaling
laws in biology (West, Brown & Enquist, 1997).
Box dimension is just one measure of the intricate-
ness of fractals. Other denitions of fractal dimension
have been proposed in the literature (Mandelbrot
1977, 1983; Falconer, 1985; Barnsley, 1988; Feder,
1988), some better than others at singling out certain
features, or spatial correlations. However, they do
not give much insight into how fractals come about
in nature. Some mathematical algorithms have been
proposed to construct such fractals as the Julia and
Mandelbrot sets or even fractal landscapes, but the
results usually lack the richness and depth of the true
fractals observed in the world around us.
(3) Fractals in time: 1/f-noise
In this subsection, we will introduce the notion of
icker or 1\f-noise, which is considered one of the
footprints of complexity. We will also see how it
diers from white or Brownian noise using the
spectral analysis method (Section II.3.a), Hursts
rescaled range analysis (Section II.3.b) and the
iterated function system method (Section II.3.c).
Let us consider a record in time of a given
quantity h of a system. h can be anything from a
temperature, a sound intensity, the number of species
in an ecosystem or the voltage at the surface of a
neuron. Such a record is obtained by measuring h at
discrete times t
!
,t
"
,t
#
, ...,t
N
, giving a series of data
ot
i
,h(t
i
)q, i l1,..., N. This time series, also called
signal or noise, can be visualized by plotting h(t) as
a function of t.
Fig. 3 shows three types of signals h(t) which will
be of interest to us (this subsection merely reproduces
the discussion from Press, 1978): white noise, icker-
or 1\f-noise and Brownian noise. Fig. 3A represents
what is usually called white noise: a random
superposition of waves over a wide range of
frequencies. It can be interpreted as a completely
168 T. Gisiger
White noise
A
B
1/fnoise
Brownian noise
20
15
10
5
0
5
0 100 200 300 400 500
Time t
S
i
g
n
a
l
h
(
t
)
C
2
0
2
2
0
2
Fig. 3. Three examples of signals h(t) plotted as functions
of time t : white noise (A), 1\f- or icker noise (B) and
Brownian noise (C).
uncorrelated signal : the value of h at some time t is
totally independent of its value at any other instant.
An example is the result of tossing a coin N
consecutive times and recording the outcome each
time. Fig. 3A shows an example of white noise that
was obtained using, instead of a coin, a random
number generator with a Gaussian distribution (see
for instance Press et al., 1988). This gives a signal
which stays most of the time close to zero, with rare
and punctual excursions to higher values.
Fig. 3C represents Brownian noise, so called
because it resembles the Brownian motion of a
particle in one dimension: h(t) is then the position of
the particle as a function of time. Brownian motion
of a particle in a uid is created by the random
impact of the liquids molecules on the immersed
particle, which gives the latter erratic displacement.
This can be reproduced by what is called a random
walk as follows : the position h of the particle at some
time tj1 is obtained by adding to its previous
position (at time t) a random number (usually
drawn using a Gaussian distribution) representing
the thermal eect of the uid on the particle. The
signal h obtained is therefore strongly correlated in
time as the particle remembers well where it was a
few steps ago. We see that the curve wiggles less than
that of white noise, and that it makes large excursions
away from zero.
The curve in Fig. 3B is dierent from the rst two
but it shares some of their characteristics. It has a
tendency towards large variations like the Brownian
motion, but it also exhibits high frequencies like
white noise. This type of signal seems then to lie
somewhere between the two, and is called icker or
1\f-noise. It is this type of signal which will interest
us in the present review because it exhibits long
trends which can be interpreted as the presence of
memory, an interesting feature in biological systems
[the method presented in Press (1978) was used to
obtain the sample shown in Fig. 3B]. Such signals
have been observed in many phenomena in physics
(see Press, 1978 and references therein): light
emission intensity curves in quasars, conduction of
electronic devices, velocities of underwater sea
currents, and even the ow of sand in an hour glass
(Schick & Verveen, 1974). It is also present in some
of the biological systems presented in this review, as
well as in other phenomena such as the cyclic insulin
needs of diabetics (Campbell &Jones, 1972), healthy
heart rate (Pilgram & Kaplan, 1999), Physarum
polycephalum streaming (Coggin & Pazun, 1996) and
in some aspects of rat behaviour (Kafetzopoulos,
Gouskos & Evangelou, 1997), to name only a few.
Like fractals, icker noise can be produced math-
ematically in several ways [see Mandelbrot (1983)
and Press (1978) for instance] though these algo-
rithms do not really help understand how it comes
about in nature.
Next, we describe mathematical methods which
can distinguish icker noise from random or Brow-
nian noise (for more details on these methods and an
example of their application to the investigation of
insect populations, see Miramontes & Rohani,
1998).
(a) Power spectrum of signals
The power spectrum P( f ) of a signal h(t) is dened
as the contribution of each frequency f to the signal
h(t). This is the mathematical equivalent of a
spectrometer analysis, which decomposes a light
beam into its components in order to evaluate their
relative importance (see for instance Press et al., 1988
for a denition of the power spectrum of a signal).
169 Scale invariance in biology
White noise
A
B
1/fnoise
Brownian noise
10
0
Frequency f
P
o
w
e
r
s
p
e
c
t
r
u
m
P
(
f
)
10
1
10
2
10
3
10
0
10
1
10
2
10
3
10
2
10
0
10
2
10
2
10
1
C
Fig. 4. Power spectrum of the signals shown in Fig. 3 of
white (A), 1\f- (B) and Brownian noise (C). The lines
tted to the spectra have slopes l0.01, lk1.31 and
lk1.9, respectively.
We present in Fig. 4 the power spectra of the signals
of Fig. 3.
By analogy with white light, which is a super-
position of light of every wavelength, white noise
should have a spectrum with equal power P( f ) at
every frequency f. This is indeed what Fig. 4A
shows. P(f) can be expressed as :
P( f ) `f
(6)
where is the gradient of the line tted to the
spectrum in Fig. 4. We nd l0.01, consistent with
P( f ) `f !.
The power spectrum of a Brownian signal also
follows a straight line on a loglog plot with a slope
equal to k2 (the line tted to our data in Fig. 4C
gives lk1.9, in reasonable accord with P( f ) `
1\f #). The power spectrum of a signal gives a
quantitative measure of the importance of each
frequency. For the Brownian motion, P( f ) goes
quickly to zero when f goes to innity, illustrating
why h(t) wiggles very little: the signal has a small
content in high frequencies. The large oscillations,
which correspond to low frequencies, constitute a
large part of the signal. Dominance of these low
frequencies can be viewed as the persistence of
information in the random walk mentioned earlier.
Flicker noise, or 1\f-noise, is dened by the power
spectrum:
P( f ) `
1
f
(7)
or more generally, as in equation (6) with .
[k1.5,k0.5]. The line tted to the data of Fig. 4B
has a slope lk1.31, well in the right range. The
interest in icker noise is motivated by its strong
content in both small and large frequencies. Be-
having roughly like 1\f, P( f ) diverges as f goes to
zero, which suggests, as in the case of Brownian
motion, long-time correlations (or memory) in the
signal. But, in addition, P( f ) goes to zero very slowly
as f become large and the accumulated power stored
in the high frequencies is actually innite. Any
spectrum with roughly equal to k1 will have these
two characteristics, which explains the somewhat
loose denition of 1\f-noise. Flicker noise is therefore
a signal with a power spectrum without any
characteristic frequency or, equivalently, time scale:
this is reminiscent of the notion of fractals, but in
space-time instead of just space.
(b) Hursts rescaled range analysis method
By the 1950s, H. E. Hurst (Hurst, 1951; Hurst,
Black & Samayka, 1965) had started the investi-
gation of long- and medium-range correlations in
time records. He made his life work the study of the
Nile and the management of its water. To help him
build the perfect reservoir, one which would never
overow or go dry, he developed the rescale range
analysis, a method, he found, that detects long-range
correlations in time series (this observation was later
put on more solid theoretical grounds by Mandel-
brot : see Mandelbrot, 1983). Using this method, he
could measure the impact of past rainfall on the level
of water. For a very thorough review of this method,
see Feder (1988).
This method associates to a time series an
exponent H which takes its values between 0 and 1.
If H1\2, the signal h(t) is said to exhibit
170 T. Gisiger
persistence: if the signal has been increasing during
a period prior to time t, then it will show a
tendency to continue that trend for a period after
time t. The same is true if h(t) has been decreasing:
it will probably continue doing so. The signal will
then tend to make long excursions upwards or
downwards. Persistence therefore favours the pres-
ence of low frequencies in the signal. This is the case
for the random walk of Fig. 3C, for which we nd H
0.96. As H goes towards 1, the signal becomes
more and more monotonous. The 1\f-noise of Fig.
3B has a Hurst exponent H0.88 which is a further
indication of long-term correlations in the signal.
When H1\2, h(t) is said to exhibit anti-
persistence: whatever trend has been observed prior
to a certain time t for a duration , will have a
tendency to be reversed for the following period .
This suppresses long-term correlations and favours
the presence of high frequencies in the power
spectrum of the signal. The extreme case where H
0 is when h(t) oscillates in a very dense manner.
The case in between persistence and anti-per-
sistence, Hl1\2, is when there is no correlation
whatsoever in the signal. This is the case of the white
noise of Fig. 3A for which we nd H0.57,
consistent with the theoretical value of Hl1\2.
It can be shown that the dierent signals h(t)
which we have presented here (white, icker and
Brownian noises) are curves which ll space
[spanned by the time and h(t) axes] to a certain
extent : they are in fact fractals and their box
dimension can be expressed as a function of the
Hurst exponent H (see Mandelbrot, 1983 and Feder,
1988 for details).
(c) Iterated function system method
Another way of dierentiating time series with long-
range correlations from ordinary noise makes use of
the iterated function systems (IFS) algorithm intro-
duced by Barnsley (1988). This very ingenious
procedure was rst proposed by Jerey (1990) to
extract correlations in sequences of genes from DNA
strands.
The method associates to a time series a two-
dimensional trajectory, or more accurately a cloud
of points, which gives visual indications on the
structure and correlations in the signal (see Jerey,
1990; Miramontes & Rohani, 1998 for details and
Mata-Toledo & Willis, 1997). The main strength of
this method rests in that it does not require as many
data points as the power spectrum or the rescaled
range analysis methods.
A
B
1
0
0
C
1
0
1
0 1
0
1
1
0
Fig. 5. Results of the iterated function system (IFS)
method when applied to the time series shown in Fig. 3 of
white noise (A), icker noise (B) and Brownian signal (C).
Fig. 5 shows the trajectories associated with the
signals of Fig. 3. The method gives dramatically
dierent results for the three signals. The trajectory
associated with white noise lls the unit square with
dots without exhibiting any particular pattern (Fig.
5A). On the other hand, the Brownian motion gives
a trajectory that spends most of its time near the
edges of the square (Fig. 5C). This is due to the
large, slow excursions of the signal, which produced
the 1\f # dependence of the power spectrum: the
171 Scale invariance in biology
signal h(t) stays in the same size range for quite some
time before moving to another. This produces a
trajectory that aggregates near corners, sides and
diagonals. When the signal nally migrates to
another interval, it will produce a trajectory which
follows the edges or the diagonals of the square.
Things are quite dierent for the icker noise (Fig.
5B). The pattern exhibits a complex structure which
repeats itself at several scales, and actually looks
fractal. The divergence of the power spectrum at low
frequencies makes the trajectory spend a lot of time
near the diagonals and the edges, similarly to the
case of the Brownian motion. However, enough high
frequencies are present to move the dot away from
the corners for short periods of time. The result is the
appearance of patterns away from the edges and the
diagonals. Their regular structure and scaling
properties make them easy to recognize visually even
in short time series.
(4) Power laws in physics: phase transitions
and universality
In this section, I briey introduce the notion of
critical phenomena. This is relevant to the general
subject of this review because systems at, or near,
their critical point exhibit power laws and fractal
structures. They also illustrate the very powerful
notion of universality, which is of great interest to the
study of complex systems in physics and biology.
Critical systems being an active and vast eld of
research in physics, it is not the goal of the present
review to give it a complete introduction. I will only
justify the following armations [being obviously
rather technical, Sections II.4.a and II.4.b can be
skipped on rst reading]:
(1) Systems near a phase transition become
critical : they do not exhibit any characteristic length
scale and spontaneously organize themselves in
fractals (see Section II.4.a).
(2) Critical systems behave in a simple manner:
they obey a series of power laws with various
exponents, called critical exponents , which can be
measured experimentally (see Section II.4.a).
(3) Experiments during the 1970s and 1980s
showed that critical exponents of materials only
come with certain special values : a classication of
substances can therefore be developed, where materi-
als with identical exponents are grouped together in
classes. The principle claiming that all systems
undergoing phase transitions fall into one of a
limited set of classes is known as universality (see
Section II.4.b).
This will be sucient to the needs of this review
where we will apply these concepts to biological
systems. For a more detailed account of critical
phenomena, the reader is referred to Wilson (1979)
and the very abundant literature on the subject (see
for instance Maris & Kadano, 1978; Le Bellac,
1988; Biney et al., 1992).
(a) Critical systems, critical exponents and fractals
Everybody is familiar with phase transitions such as
water turning into ice. This is an example of
discontinuous phase transition: matter suddenly goes
froma disordered state (water phase) to an organised
state (ice phase). The sudden change in the
arrangement of the water molecules is accompanied
by the release of latent heat by the system. Here, we
will be interested in somewhat dierent systems.
They still make transitions between two dierent
phases as the temperature changes, but they do so in
a smooth and continuous manner (i.e. without
releasing latent heat). These are called continuous
phase transitions and they are of great experimental
and theoretical interest. (Here, I use the terms
discontinuous and continuous for phase transi-
tions ; physicists prefer the terms rst order and
second order phase transitions.)
The classical example of a system exhibiting a
continuous phase transition is a ferromagnetic
material such as iron. Water too can be brought to
a point where it goes through a continuous phase
transition: it is known as the critical point of water,
characterized by the critical temperature T
c
l
647 K and the critical density
c
l0.323 g cm
$.
There, the liquid and vapour states of water coexist,
and in fact look alike. This makes the dynamics of
the system dicult, even confusing, to describe with
words. Here, we will therefore use the paradigm of
the ferromagnetic material as an example of critical
system, i.e. of a system near a continuous phase
transition.
It was shown by P. Curie at the turn of the
century that a magnetized piece of iron loses its
magnetization when it is heated over the critical
temperature T
c
1043 K. Similarly, if the same
piece is cooled again to below T
c
, it spontaneously
becomes remagnetized. This is an example of a
continuous phase transition because the magnetiz-
ation of the system, which we represent by the vector
M, varies smoothly as a function of temperature:
Ml0 in the unmagnetized phase and it takes a non-
zero value in the magnetized phase. The magnetiz-
ation M of the sample, which is easily measured
172 T. Gisiger
Fig. 6. Illustration of a two-dimensional sample of
material in the magnetized phase (TT
c
): the spins m
of the atoms, represented by the small arrows, are all
aligned and produce a non-zero net magnetization M `
sample
m for the sample. T, temperature; T
c
, critical
temperature.
using a compass or more accurately a magnetometer,
therefore allows us to determine which phase the
piece of iron is in at a given instant.
To understand better the physics at work in the
system, we have to look at the microscopic level. The
sample is made of iron atoms arranged in a roughly
regular lattice. In each atom, there are electrons
spinning around a nucleus. This creates near each
atom a small magnetic eld m called a spin, which
can be approximated roughly by a little magnet like
a compass arrow (see Fig. 6 for an illustration). m is
of xed length but it can be oriented in any direction.
The total magnetization of the material is pro-
portional to the sum of the spins :
M`
sample
m (8)
over all the atoms of the sample. Therefore, if all the
spins point in the same direction, their magnetic
eects add up and give the iron sample a non-zero
magnetization M. If they point in randomdirections,
they cancel each other and the sample has no
magnetization (Ml0). This microscopic picture
can be used to explain what happens during the
phase transition. If we start with a magnetized block
of iron (all mpointing roughly in the same direction)
and heat it up, the thermal agitation in the solid will
disrupt the alignment of the spins, therefore lowering
the magnetization of the sample. When TlT
c
, the
agitation is strong enough to completely destroy the
alignment and the total magnetization is zero at T
c
and for any temperature larger than T
c
. Similarly,
when the sample is hot and is then cooled to below
the critical temperature, the spins spontaneously
align with each other. It can be shown that for T
slightly smaller than T
c
, the magnetization Mfollows
a power-law function of the temperature:
QMQ `QT
c
kTQ
, (9)
where can be measured to be approximately
0.37 (if T is larger or equal to T
c
then Ml0).
therefore quanties the behaviour of the magnetiz-
ation of the sample as a function of temperature: it
is called a critical exponent. It can be shown that
other measurable quantities of the system, such as
the correlation length dened below, obey similar
power laws near the critical point (but with dierent
critical exponents : besides , ve other exponents
are necessary to describe systems near phase tran-
sitions) (Maris & Kadano, 1978; Le Bellac, 1988;
Biney et al., 1992).
Let us consider the sample at a given temperature,
and measure the eect that ipping one spin has on
the other spins of the system. If TT
c
, the majority
of spins will be aligned in one direction. Flipping one
spin will not inuence the others because they will all
be subject at the same time to the much larger
magnetic eld of the rest of the sample. If TT
c
,
changing the orientation of one spin will modify only
that of its neighbours since the net magnetization of
the material is zero. However, near the phase
transition (TT
c
), one spin ip can change the
spins of all the others. This is because as we approach
the critical point, the range of interactions between
spins gets innite: every spin interacts with all other
spins. This can be formalized by the correlation
length, , dened as the distance at which spins
interact with each other. Near the phase transition,
it follows the power law:
`QTkT
c
Q
, (10)
with 0.69 for iron, and therefore diverges to
innity as T goes to T
c
: there is no characteristic
length scale in the system. Another way of under-
standing this phenomenon is that, as T is ne-tuned
to T
c
, the spins of the sample behave like a row of
dominoes where the fall of one brings down all the
others. Here also, the interaction of one domino
extends eectively to the whole system. This seems to
take place by the spins arranging themselves in a
scale-free, i.e. fractal, way (see Fig. 7). Fractal
structures have been conrmed both experimentally
and theoretically: at TlT
c
, the spins are arranged
173 Scale invariance in biology
A
B
C
Fig. 7. Spin disposition of a sample as simulated by the
Ising model. Each black square represents a spin pointing
up, and the white ones stand for a spin pointing down. (A)
TT
c
: almost all the spins are pointing in the same
direction, giving the sample non-zero magnetization
(magnetized phase). (B) TlT
c
: at the phase transition,
the net magnetization Ml0 but the spins have arranged
themselves in islands within islands of spins pointing in
opposite directions, which is a fractal pattern. (C) T
T
c
: the system has zero magnetization and only short-
range correlations between spins exist (unmagnetized
phase). T, temperature; T
c
, critical temperature.
in islands of all sizes where all m point up, within
others where all point down, and so on at smaller
scales, with net magnetization which is zero [see Fig.
7 for a computer simulation (Gould & Tobochnik,
1988) of a particularly simple spin system: the Ising
model].
(b) Universality
During the 1970s and 1980s, there was a great deal
of experimental work performed to measure the
critical exponents of materials : polymers, metals,
alloys, uids, gases, etc. It was expected that to each
material would correspond a dierent set of ex-
ponents. However, experiments proved this sup-
position wrong. Instead materials, even those with
no obvious similarities, seemed to group themselves
into classes characterised by a single set of critical
exponents. For instance, it can be shown that when
taking into account experimental errors, one-dimen-
sional samples of the gas Xe and the alloy -brass
have the same values for critical exponents (see
Maris & Kadano, 1978; Le Bellac, 1988; Biney et
al., 1992). This is also the case for the binary uid
mixture of methanol-hexane, trimethylpentane and
nitroethane. This gas, alloy and liquid mixture
therefore all fall into a class of substances labeled by
a single set of critical exponents. By contrast, a three-
dimensional sample of Fe does not belong to this
class. However, it has the same critical exponents as
Ni. They therefore both belong to another class.
Since critical exponents completely describe the
dynamics of a system near a continuous phase
transition, the fact that the classication mentioned
above exists proves that arbitrary critical behaviour
is not possible. Rather, only a limited number of
behaviours exist in nature, which are said to be
universal, and dene disjoint classes called uni-
versality classes. The principle which states this
classication is therefore called universality. The
following theoretical explanation of this astonishing
fact has been proposed (Wilson, 1979; see also Maris
& Kadano, 1978; Le Bellac, 1988; Biney et al.,
1992): near a continuous phase transition, a given
system is not very sensitive to the nature of the
particles it is constituted of, or to the details of the
interactions which exist between them. Instead, it
depends on other, more fundamental, characteristics
of the system such as the number of dimensions of the
sample (see Wilson, 1979). It is a point of view which
ts well within the philosophy of complex systems we
mentioned in the introduction.
Universality has been described as a physicists
dream come true. Indeed, what it tells us is that a
system, whether it is a sample in a laboratory or a
mathematical model, is very insensitive to details of
its dynamics or structure near critical points. From a
theoretical point of view, to study a given physical
system, one only has to consider the simplest
mathematical model possibly conceivable in the
same universality class. It will then yield the same
critical exponents as the system under study. A
famous example is the Ising model proposed to
explain the ferromagnetic phase transition and
which we introduce now. It represents spins of the
iron atoms by a binary variable S which can either
be equal to 1 (spin up) or to k1 (spin down). The
174 T. Gisiger
spins are distributed on a lattice and they interact
only with their nearest neighbours. Even though this
model simply represents spins as j or k, does not
allow for impurities, irregularities in the disposition
of spins, vibrations, etc., it yields the right critical
exponents.
Critical phenomena is a eld where the intuitive
idea that the description of a system is dependent on
the amount of detail put into it does not hold. As
long as a system is known to be critical, the simplest
model (sometimes simple enough to be solved by
hand) will do. This approach is not restricted to
physical systems. In fact, most of the biological
systems presented in this review will be studied using
extremely simple, usually critical, models.
III. GENERALITIES ON MODELS AND THEIR
PROPERTIES
So far, we have briey introduced the notion of
complex systems and why they are of interest in
science. We have also presented in detail the
property of some systems which do not possess any
characteristic scale, and how it can be observed:
scale invariance in fractals (measurable by the box
counting method), correlations on all time scales in
1\f-noise (diagnosed by the power spectrum of the
signal, for instance), and power-law distribution of
event size or duration. Finally, we have briey
described physical critical systems which exhibit
scale-free behaviour naturally when one tunes the
temperature near its critical value. As a bonus, we
encountered the notion of universality which tells us
that critical systems may be accurately described by
models which only approximate roughly the interac-
tions between its constituents.
We therefore have the necessary tools to detect
scale invariance in biological systems, and a general
principle, criticality, which might explain how this
scale-free dynamics arises. However, to make contact
between our understanding of a system and ex-
perimental data, one needs mathematical models.
Major types of models used in biology are dierential
equation systems, iterative maps and cellular auto-
mata, to name only a few. Here, we will review in
turn those which can produce power laws and scale
invariance.
We start in Section III.2 with dierential equa-
tions and discrete maps which exhibit transitions to
chaotic behaviour. Section III.3 presents spatially
discretized models like percolating systems and
cellular automata. We then move on, in Section
III.4, to the concept of self-organized criticality,
illustrating it with the now famous sandpile model.
This will be done with special care since most of the
models presented in this review are of this type. At
the end of this Section we take a few steps back from
these developments to discuss from a critical point of
viewthe limitations of the complex systems paradigm
(Section III.5).
(1) Generalities
Changeux (1993) gives the following denition for
theoretical models : In short, a model is an
expanded and explicit expression of a concept. It is
a formal representation of a natural object, process or
phenomenon, written in an explicit language, in a
coherent, non contradictory, minimal form, and, if
posssible, in mathematical terms. To be useful, a
model must be formulated in such a way that it
allows comparison with biological reality. [] A
mathematical algorithm cannot be identied with
physical reality. At most, it may adequately describe
some of its features
I agree with this denition. However, I feel that
two points need clarication.
First, there is the issue of the relationship between
a model and experimental data. In order to be
useful, a model should always reproduce, to a
sucient extent, the available measurements. It is
from an understanding of these data that hypotheses
about the system under study can emerge and be
crystalized into a model. The latter can then be
tested against reality by its ability to account for the
experimental data and make further predictions.
Without this two-way relationship, the line between
theoretical construction and theoretical speculation
can become blurred, and easily be crossed.
History has shown that, in physics for instance (see
Einstein & Infeld, 1938 for a general discussion on
modelling and numerous examples), some of the
most signicant theoretical advances were made by
the introduction of powerful new concepts to
interpret a mounting body of experimental evidence.
However, by no means do I imply here that
theoretical reection should be conned to the small
group of problems solvable in the short of medium
term. I just want to stress that one should be careful
about the validity and robustness of results obtained
using models based on intuition alone or sparse
experimental ndings.
Second, I think that the adjective minimal
deserves to be elaborated on somewhat (see also Bak,
1996).
175 Scale invariance in biology
A good starting point for our discussion is weather
forecasting. The goal here is to use mathematical
equations to predict, from a set of initial conditions
(temperature, pressure, humidity, etc.), the state of
the system (temperature, form and amount of
precipitation, etc.) with the best precision possible
over a range of approximately 1236 hours. In this
case, minimal translates to a model made of
hundreds or thousands of partial dierential equa-
tions with about the same amount of variables and
initial conditions. Such complexity allows the in-
clusion of a lot of details in the system, and is often
considered to be the most realistic representation of
nature. However, this approach has two serious
limitations. The rst is that the equations used are
often non-linear and their solutions are unstable and
sensitive to errors in the initial conditions : any small
error in temperature or pressure measurements at
some initial time will grow and corrupt the predic-
tions of the model as the simulation time increases.
This is the well-known buttery eect (Lorenz,
1963) which prevents any long-term weather fore-
casting. It also explains why such models can give
little insight into global warming or the prediction of
the next ice age. The second limitation of this
approach stems from the sheer complexity of the
model, which usually needs supercomputers to run
on. It makes it dicult to get an intuitive feel of the
dynamics of the system under investigation. One
then has a right to wonder what has been gained,
besides predictive power: prediction is then not
always synonymous with understanding.
At the other end of the complexity spectrum stand
models which have been simplied so much that
they have been reduced to their backbone, taking
the minimal adjective as far as one dares. Such
models are sometimes frowned upon by experimen-
talists, who do not recognize in them the realization
of the system they study. Theorists, on the other
hand, enjoy their simplicity: such models yield more
easily to analysis and can be simulated without
resorting to complicated algorithms or supercom-
puters. They too can be extremely sensitive to
variations in initial conditions (like the Lorenz
model presented in Section III.2) but in a more
tractable and controled way. A lesser predictive
power, sometimes only qualitative predictions, is
usually the price to pay for this simplicity. However,
as we saw, in certain cases like the critical systems
described in Section II.4, simplicity can coexist with
spectacularly accurate quantitative results. One only
has to choose a simple, or simplistic, model in the
same universality class as the system under study. In
20
0
20
40
30
20
10
0
10
0
10
x
z
y
Fig. 8. Classical example of chaotic behaviour generated
by the equations of Lorenz (1963). The trajectory winds
forever in a nite portion of space without ever inter-
secting itself, creating a structure called a strange
attractor. It is fractal with box dimension 2.05, zero
volume but innite area.
this review, we will consider only models belonging
to this latter class.
(2) Chaos: iterative maps and dierential
equations
While reading popular literature, one can get the
impression that chaos and fractals are two sides of
the same coin. As we saw, fractals were introduced
by Mandelbrot (1977, 1983) in the 1970s. At about
the same time, Lorenz (1963) was discovering that
the simple set of non-linear dierential equations he
had devised as a toy model for weather forecasting
exhibited strange, unheard-of behaviour. First, the
trajectory solution to these equations winds in an
unusual, non-periodic way (see Fig. 8): it has
actually been shown to follow a structure with box
dimension 2.05 (see Strogatz, 1994) which is
therefore fractal. Lorenz coined the now-famous
term strange attractor to designate it. [Attractors
can be dened as the part of space which attracts the
trajectory of a system. It can be anything from a
point (when the system tends towards an equi-
librium), a cycle (for periodic motion), to a fractal
structure. The latter is then called a strange
176 T. Gisiger
attractor.] Second, this trajectory is also extremely
sensitive to changes in initial conditions. Any small
modication
!
will grow with time t as e
t
, with ,
called a Lyapunov exponent, having a value of
roughly 0.9 for the system of Lorenz (Lorenz, 1963;
May, 1976; Feigenbaum, 1978, 1979; Strogatz,
1994). The variation induced by the introduction of
!
will then increase extremely quickly: this model is
exponentially sensitive to changes in its initial
conditions. This makes any long-term weather
forecasting with such systems impossible since any
error in the initial conditions (which are present in
any measurements) will corrupt its predictions.
The word chaos was chosen to describe the
dynamics of systems which do not exhibit any
periodicity in their behaviour and are exponentially
sensitive to change in their initial conditions. [Thus,
they have positive values of their Lyapunov ex-
ponents. For negative values of , e
t
will quickly
go to zero and changes in initial conditions will
therefore not aect the dynamics.] This behaviour
was put in a wider context by the work of
Feigenbaum (1978, 1979) who showed that, by
adjusting the values of their parameters, certain non-
linear models can shift from a non-chaotic regime
(i.e. periodic and not very sensitive to variations in
their initial conditions) to a chaotic state (i.e. non-
periodic with high sensitivity to initial conditions).
This transition takes place through a succession of
discrete changes in the dynamics of the system,
which he called bifurcations . Feigenbaum (1978,
1979) used, in his studies a simplied model of an
ecological system, called the logistic map, proposed
earlier by May (1976).
This classical model is a simple, non-linear,
iterative equation which describes the evolution of
the population x
t
of an ecosystem as a function of
time t. Though comprising only one parameter r,
which takes its values between 0 and 4 and quanties
the reproductive capabilities of the organisms, this
model is capable of producing time series x
"
, x
#
, , x
N
with increasingly complicated structures as r grows
from small values to larger ones. For r close to zero,
the population of the ecosystem tends to stabilize
with time to a constant value. For slightly higher r,
it becomes periodic, oscillating between two values
for the population. This radical change in dynamics
is an example of bifurcation. As r rises further, the
time series of the population becomes more and more
complicated as new values are added to the cycle by
further bifurcations. For the value r r
_
)
10
1
10
0
10
0
10
1
10
2
A
B
Fig. 10. (A) Signal for the Manneville iteration map
tuned at the edge of chaos (upper), with the same signal
after being ltered (lower; a spike was drawn each time
the signal went over the value 0.6810). (B) Plot of the
distribution D() of time delays between consecutive
spikes.
chaotic systems will do in the long run. There is a loss
of information as time passes. This suggests that, by
looking at this from the opposite point of view, a
chaotic system has a very short memory: it does not
remember where it was for very long. It might
therefore not be a good candidate to describe
biological systems which must adapt and learn all
the time.
However, as Bak (1996) mentions, the map
exhibits a far richer behaviour at r lr
_
, right at the
point between periodic behaviour and the chaotic
regime, therefore sometimes called the edge of
chaos . Fig. 9A shows the power spectrum of the
signal there: the structure is certainly not 1\f, but
it is interesting and appears to be self-similar. It
can also be shown that the time series with innite
periodicity produced for r lr
_
falls on a discrete
fractal set with a box dimension smaller than 1,
called a Cantor set. Furthermore, Costa et al. (1997)
have shown that at the edge of chaos, the sensitivity
to initial conditions of the logistic map is a lot milder
than in the chaotic regime. There, the Liapunov
exponent is zero and instead the sensitivity to
178 T. Gisiger
changes in initial conditions follows a power law
which, as we saw in Section II.1, is a lot milder than
the exponential e
t
. This will sustain information in
the system a lot longer, and might provide long-term
correlation in the time evolution of the population x
t
.
It was, however, Manneville (1980) who rst
showed that, tuned exactly, an iterative function
similar to that of May (1976) can produce interesting
behaviour and power laws (see also Procaccia &
Schuster, 1983; Aizawa, Murakami & Kohyama,
1984; Kaneko, 1989). Fig. 10A (upper panel) shows
the signal produced by his map: it is formed by a
succession of peaks which, when interpreted as a
series of spikes (Fig. 10A, lower panel) looks similar
to the train of action potentials measured at the
membrane of neurons (we will return to this analogy
in Section VI.3). Fig. 10B shows the distribution of
intervals between successive spikes : it follows a
power law D() `
"
n
$ over several orders of
magnitude. Manneville (1980) also computed the
power spectrum of the signal and found that is was
1\f. This shows that it is indeed possible to generate
power laws and 1\f-noise from simple iterative maps
by ne-tuning the parameters of the system at the
edge of chaos (see Manneville, 1980 for details).
(3) Discrete systems in space: cellular
automata and percolation
In this Section, we will introduce percolation systems
and cellular automata. This will be helpful since
most models in this review belong to one of these two
types.
Iterative functions and dierential equations are
not always the most practical way to describe
biological systems. Let us take the following example.
How should we go about building a model for tree or
plant birth and growth in an empty eld? The
model should be able to predict how many plants
will have grown after a given time interval, and also
if they will be scattered in small patches or in just one
large patch which spans the whole eld. We will
make the simplifying assumption that seeds and
spores are carried by the wind over the eld, and
scattered on it uniformly.
This is of course a dicult problem, and using
dierential equations to solve it might not be the
most practical method. The underlying dynamics of
how a seed lands on the ground and decides, or not,
to produce a plant is already delicate. Further, it
depends also on the types of seed and ground
involved, the amount of precipitation, the vegetation
A B C
Fig. 11. Simulation of plant growth in a square empty
eld divided into Nl2500 small areas. Cells with a plant
on them are represented by a black square. Free areas are
symbolized by white squares. (A) p l0.1, (B) p l0.3,
and (C) p l0.58, where p is the probability that a plant
grows on each small area.
already present, etc. A more practical approach
might be a probabilistic one.
Here, we rst divide the eld into N small areas,
using a square lattice for instance. Each cell must be
small enough that at most one plant can grow on it
at a time. We then dene a number p . [0, 1] which
represents the probability that a plant grows on each
small area. An approximate value for p can be found
experimentally by reproducing the conditions found
on location. We then proceed to ll all the cells
using the probability p. Fig. 11 shows an example for
p l0.1, 0.3 and 0.58. We notice, as expected, that
the number of plants present and the size of the
patches increases with p.
Repeating the simulation several times also gives
the following interesting result : the proportion of
simulations where the eld is spanned from border to
border by a single patch varies abruptly as a
function of the probability p. It is close to zero if p
is smaller than approximately 0.59, and almost
equal to 1 if p exceeds this value. Indeed, it has been
shown that such systems, which are called perco-
lation systems because of their similarity to the
percolation of a uid in a porous medium (for
reviews see Broadbent & Hammersley, 1957;
Stauer, 1979; Hammersley, 1983), become critical
when p is ne-tuned to the value p
c
0.59275: the
shape of the patch spanning the eld is fractal, and
the properties of the systemobey power law functions
of pkp
c
. The probability p therefore plays a role
similar to temperature T in the systems of Section
II.4.
How will this model help us answer the questions
asked at the beginning of this Section? The number
of plants which will grow on the eld depends on p
and the model predicts the value Np. As to whether
plants will group themselves into small patches or
spread all over the eld, the answer depends also on
179 Scale invariance in biology
p (which one should be able to estimate exper-
imentally). As we just saw, if p 0.6 then small
scattered patches are probable. Otherwise, complete
colonisation of the eld by the plants is almost
certain.
This very simple model can be generalized to a
larger number of dimensions instead of just the two
considered here. For instance, one can easily con-
ceive of percolation in a three-dimensional cube.
Dynamics in a number of dimensions larger than
three, though harder to represent mentally, is easily
implementable on computers.
This model will be generalized in later sections to
implement the dynamics of rain forests, forest res
and epidemics. Indeed, more complex dynamics can
easily be developed. For instance, we can change the
rules according to which we ll the areas from the
eld. So far, we have only used a random number
generator for each cell. This means that there are no
interactions between plants : the growth at one place
does not aect the growth at neighbouring sites.
Overcrowding is therefore not present in our model.
This can easily be remedied using slightly more
complicated rules.
A good example of such models is the so-called
Game of Life introduced by Conway (see Gardner,
1970 and Berlekamp, Conway & Guy, 1982) in the
1970s, which mimics some aspects of life and
ecosystems. The Game of Life is dened on a square
lattice. Each lattice site can be either occupied by a
live being or be empty. Time evolves in a discrete
fashion. At each time step, all the sites from the
lattice are updated once using the following rules :
(1) An individual will die of over-exposure if it
has less than two neighbours, and of over-crowding
if it has more than three neighbours. The site it
occupied previously will then become empty at the
next time step.
(2) A new individual will be born at an empty site
only if this site is surrounded by three live neigh-
bours.
The dynamics generated by these simple rules has
proven to be very rich, with structures which glide
without changing shape, others that do not move at
all but are unstable, etc. There has been a recent
revival of interest in this model as power laws and
fractals appear in its dynamics, possibly hinting at
the presence of criticality in models of ecosystems
(see Bak, Chen & Creutz, 1989 and Sales, 1993).
Mathematical models of this latter type, called
cellular automata (Wolfram, 1983, 1984, 1986; see
also Langton, 1990) allow a simple and ecient
implementation of complicated interactions between
the constituents of a system using rules such as those
above.
(4) Self-organized criticality: the sandpile
paradigm
Here, I present the principle of self-organized
criticality introduced by Bak, Tang and Wiesenfeld
(1987) and its paradigm, the sandpile. This subject
has attracted quite a lot of attention since its
introduction: hundreds of articles are published on it
every year in physics, mathematics, computer sci-
ence, biology and economics journals. Short reviews
are available to the interested reader wishing to
learn more about this fast-moving eld (Bak &
Chen, 1989, 1991; Bak, 1990; Bak & Paczuski, 1993,
1995; Bak, 1996).
Self-organized criticality is a principle which
governs the dynamics of systems, leading them to a
complex state characterized by the presence of frac-
tals and power-law distributions. This state, as we
will see, is critical. However, self-organized critical
systems dier from the systems we presented in
Section II.4. There, one had to ne-tune a para-
meter, the temperature T, to be in a critical state.
Here, it is the dynamics of the system itself which
leads it to a scale-free state (it is therefore self-
organized). The classical example of such systems is
the sandpile (Bak et al., 1987; Bak, 1996) which we
turn to now.
Let us consider a level surface, such as a table, and
some apparatus which allows to drop sand on it,
grain by grain, every few seconds. With time, this
will create a growing accumulation of sand in the
middle of the surface. At rst, the sand grains will
stay where they land. However, as the sand pile gets
steeper, instabilities will appear: stress builds up in
the pile but it is compensated by friction between
sand grains. After a while, stress will be high enough
that just adding one single grain will release it in the
form of an avalanche: a particle of sand makes an
unstable grain topple, and briey slide. This has the
eect of slightly reducing the slope of the sandpile,
rendering it stable again until more sand is added.
Such toppling events will be small in the beginning,
but they will grow in size with time. After a while,
the sandpile will reach a state where avalanches of
all sizes are possible: from one grain of sand, to a
large proportion of the whole pile. This state is
critical since it exhibits the same domino eect
which was present in the spin system: adding one
grain of sand might aect all the other grains from
the pile. Further, this state has been attained without
180 T. Gisiger
Fig. 12. Cellular automaton introduced by Bak et al.
(1987) on a square lattice of size Nl20 in the critical
state. Grains of sand are represented by cubes, which
cannot be piled more than three particles high. Notice the
correlations in the spatial distribution of the sand.
any ne tuning of parameters : all one has to do to
reach it is to keep adding sand slowly enough so that
avalanches are over before the next grain lands.
To better understand the dynamics of the system,
let us consider the cellular automaton proposed by
Bak et al. (1987). We represent the level surface by a
NiN square grid, and grains of sand by cubes of
unit sides which can be piled one on top of another
(see Fig. 12). Let Z(i, j) be the amount of sand at the
position (i, j) on the grid. At each time step, a
random number generator will choose at which
position (i, j) of the grid the next grain of sand will
land, or equivalently which Z will increase by one
unit :
Z(i, j) Z(i, j)j1. (11)
To model the instabilities in the sandpile, we will
not allow sand to pile higher than a certain critical
value chosen arbitrarily to be 3. If at any given time,
the height of sand Z(i, j) is larger than 3 then that
particular site will distribute four grains to its four
nearest neighbours :
Z(i, j) Z(i, j)k4 (12)
Z(ij1, j) Z(ij1, j)j1 (13)
Z(ik1, j) Z(ik1, j)j1 (14)
Z(i, jj1) Z(i, jj1)j1 (15)
Z(i, jk1) Z(i, jk1)j1 (16)
A single toppling event like equations (1216) is an
avalanche of size 1. If this event happens next to a
site where three grains are already piled up, sand
will topple there as well, giving an avalanche of size
500
400
300
200
100
0
0 20 40 60 80 100
10
3
10
2
10
1
10
0
10
0
10
1
10
2
Size s
Time t
A
v
a
l
a
n
c
h
e
s
i
z
e
s
D
i
s
t
r
i
b
u
t
i
o
n
D
(
s
)
B
A
Fig. 13. (A) Record of avalanche size as a function of time
for a 20i20 sandpile. (B) Distribution D(s) of avalanches
for the same sandpile as a function of their size s. It clearly
follows a straight line over more than two orders of
magnitude, therefore implying a power law. Fit to the
data gives D(s) `s
".!.
2, and so on. We apply the updating rule (equations
1216) until all Z are equal to 3 or smaller. The
duration of the avalanche is dened as the number of
consecutive time steps during which Z(i, j) 3
somewhere in the sandpile. The result does not look
much like a real sandpile (see Fig. 12). It looks more
like small piles of height 3 and less, held close
together. However, this automaton proves a lot
easier to implement on a computer (and runs faster)
than models of a more realistic piles, and it exhibits
the same behaviour.
We now present the results of simulations of the
cellular automaton on a 20i20 grid. Fig. 13A shows
a record of avalanche sizes as a function of time.
Notice that avalanches of a wide range of sizes occur,
and that the smaller ones are much more frequent
than the larger ones. In fact, as shown on Fig. 13B,
the distribution D(s) of avalanche size s follows the
power law D(s) `s
".!. Avalanche duration also
181 Scale invariance in biology
50
40
30
20
10
0
0 20 40 60 80 100
10
3
10
2
10
1
10
0
10
0
10
1
10
2
Time t
A
v
a
l
a
n
c
h
e
d
u
r
a
t
i
o
n
s
D
i
s
t
r
i
b
u
t
i
o
n
D
(
s
)
B
A
Duration s
Fig. 14. (A) Record of avalanche duration as a
function of time t for a 20i20 sandpile. (B) The
distribution D() of avalanches as a function of their
duration follows a straight line over roughly two orders
of magnitude therefore implying a power law. Fit to the
data gives D() `
!
n
).
has interesting properties. Fig. 14A shows the time
record of the avalanche duration for the data of
Fig. 13. The two look similar. In fact, avalanche
duration distribution also follows a power law as
shown in Fig. 14B: D() `
!
n
). We refer the
interested reader to the literature for further details
about this fascinating system (Bak, Tang & Wiesen-
feld, 1988).
It is important to mention that in this system, the
criticality is only statistical, unlike the systems near
continuous phase transitions we presented in Section
II.4. In the case of the ferromagnetic sample, when
TT
c
, changing the orientation of one spin will
aect all the other spins of the sample. It is a kind of
avalanche, but it is always as large as the whole
sample. In the sandpile, events of all size and
duration are possible but their relative probability is
given by a power-law distribution. This is strongly
reminiscent of the example of earthquakes, which we
mentioned in the introduction: there is no more
typical size for avalanches than there is a typical size
for earthquakes. In the case of the sandpile, the size
and duration of an event depends of course on where
we put the grain of sand which triggers the
avalanche, but it depends much more strongly on
the amount of stress built in the pile. At some times,
the stress will be low and small events will occur. At
other times, it will be much higher and large events
will become very probable. Apart from the number
generator which distributes sand on the pile, nothing
is random in the system. One should therefore expect
strong long-term correlations in signals produced by
the dynamics of the sandpile, which is indeed what
is observed (we will come back to this point later).
These characteristics (long-term correlation, scale
invariance, and the absence of any ne tuning)
makes self-organized criticality an attractive prin-
ciple to explain the dynamics of scale-free biological
systems.
The sandpile exhibits a few more power laws. For
instance, similarly to the logistic map at the edge of
chaos of Section III.2, the sensitivity of the pile to
changes in initial conditions is of a power-law type
(Bak, 1990). The system also exhibits fractal struc-
tures (Bak & Chen, 1989; Bak, 1990) (see also Fig.
12). It was also claimed (Bak et al., 1987, 1988; Tang
& Bak, 1988) that it emits 1\f-noise.
After the ground-breaking theoretical work of P.
Bak and his colleagues, many eorts were made to
observe experimentally self-organized criticality.
The experiments were dicult because it was
necessary, in order to measure accurately the
distributions D(s) and D(), to count all the grains of
sand which moved during an avalanche. Careful
compromises therefore had to be made while
devising experimental devices. Early eorts were not
very successful, producing puzzling, and sometimes
misleading, results. It was not until the experimental
work by Frette et al. (1996) that true power-law
scaling was measured. The pile was not made of
sand, neither was it three dimensional. Instead it was
made of rice, constrained in a two-dimensional space
between two transparent plastic pannels through
which digital cameras could, with the assistance of
computers, follow the motion of all the particles of
the pile. The power laws, signature of self-organized
criticality, were reproduced but only when using a
certain type of rice which provided enough friction
to reduce the size of avalanches. This experiment
gave insight into another power-law distribution:
that for the duration of the stay of a given grain of
rice in the pyramid (Christensen et al., 1996). It also
serves as a warning that self-organized criticality
182 T. Gisiger
might not be as general and universal a principle it
might appear at rst, since it seems to be sensitive to
details of the dynamics (such as the type of rice
used).
Another unexpected development was the dem-
onstration that the original sandpile automaton
(Bak et al., 1987) does not produce 1\f-noise, but
rather 1\f #-noise, like Brownian noise (Jensen,
Christensen & Fogedby, 1989; Kerte! sz & Kiss,
1990; Christensen, Fogedby & Jensen, 1991). How-
ever, other self-organized critical systems do give
1\f-noise (Jensen, 1990), and by incorporating
dissipation in the sandpile automaton, one can tune
the noise to be in the 1\f range (Christensen, Olami
& Bak, 1992). Finally, in a recent paper, De Los
Rios and Zhang (1999) propose a sandpile-type
model which includes dissipation and a preferred
direction for the propagation of energy which
produces 1\f-noise in systems with any number of
dimensions. Interestingly enough, because of dis-
sipation eects, this system does not exhibit power-
law distributions of spatial correlations, unlike the
original sandpile model and other self-organized
critical systems. This gives theoretical support to the
experimental observation of systems which emit 1\f-
noise but are not scale-free (see references in De Los
Rios & Zhang, 1999). More work is therefore clearly
needed to understand better the connection between
icker noise and criticality in nature.
(5) Limitations of the complex systems
paradigm
Before moving to scale invariance in biological
systems, I present a few limitations of the complex
systems paradigm proposed in the introduction, i.e.
of the study of the emergent properties of systems
once a knowledge of the interactions between their
components is accessed.
The rst main problem of this approach, which is
also present in any type of modelling, is to dene
rigorous ways in which models can be tested against
reality. In the case of scale-invariant systems, things
are relatively simple: quantities such as exponents of
power laws and fractal dimensions can be measured
experimentally and compared with predictions from
models. The case of the BelouzovZhabotinski
reaction for instance, is more complex as ways to
quantify reliably the dynamics of the system are
more dicult to nd. Indeed, it is possible to
construct models which create various spatial pat-
terns, but verifying their validity in accounting for
the mechanisms of the reaction just by comparing
by eye the structures obtained from simulations
and those observed experimentally is obviously not
satisfactory.
The second main problem is more central.
Although the idea of complex behaviours emerging
from systems of identical elementary units inter-
acting with each other in simple ways is quite
attractive, it is seldom realized in nature. Indeed,
most real systems are made of an heterogenous set of
highly specialized elements which interact with each
other in complicated ways. Genetics and cell biology,
for instance, are replete with such systems.
In the light of these arguments, we see that
although the complex system approach is an im-
portant rst step towards a unied theory of systems,
it is unlikely to be the last, and many exciting
developments will have to take place in the future.
This concludes the part of this article devoted to
mathematical notions and concepts. We can now
apply these tools to biological systems.
IV. COMPLEXITY IN ECOLOGY AND
EVOLUTION
(1) Ecology and population behaviour
(a) Red, blue and pink noises in ecology
Following the work of Feigenbaum (1978, 1979) on
the logistic map of May (1976), eorts were devoted
to analysing experimental data from animal popu-
lations using similar models. This aimed at a better
understanding of the variations in the population
density of ecosystems as well as how the environment
inuences this variable. Another question which
attracted a lot of attention was whether ecosystems
poise themselves in a chaotic regime. A considerable
amount of literature has been devoted to this subject
and the reader is referred to it for further details (see
for instance the review by Hastings et al. (1993) and
references therein). Answering this second question
is especially dicult since time series obtained from
ecosystems are usually short while methods used to
detect chaotic behaviour require a large amount of
data (Sugihara & May, 1990; Kaplan & Glass,
1992; Nychka et al., 1992): results are not always
clear cut. One could argue that, considering the
complexity of the interactions between individuals,
population dynamics should be strongly inuenced
by past events. However, as we saw earlier, one of
the main characteristics of chaotic systems is to be so
sensitive to initial conditions that information is lost
after a short period. I will not pursue this matter
183 Scale invariance in biology
further and refer the reader to Berryman & Millstein
(1989) for an interesting discussion on the subject.
The question seems to be still open.
Additional insight into the dynamics of ecosystems
was revealed by the power spectra of population
density time series from diverse ecosystems. Pimm &
Redfearn (1988) considered records of 26 terrestrial
populations (insects, birds and mammals) compiled
over more than 50 years for British farmlands.
Computing the power spectrum of these data, they
found that it contains a surprisingly high content in
low frequency, and described this signal as red
noise. This indicates the presence of slow variations
or trends in population density, which might suggest
that the ecosystem is strongly inuenced by past
events (in other words, it possesses a memory). This
came quite as a surprise since simple chaotic models,
which were thought at the time to capture the
essential ingredients of the dynamics of ecosystems,
generate time series rich in high frequency, so-called
blue signals (see Section III.2). However, it was
later shown that by including spatial degrees of
freedom, chaotic models can exhibit complex spatial
structures (such as spirals or waves), very similar to
those in the BelouzovZhabotinski experiment
(Hassel, Comins & May, 1991; Bascompte & Sole! ,
1995). With this addition, the dynamics of the model
seem to generate signals with a higher content in low
frequencies, as if the system was able to store
information in these patterns. There is, however, still
much debate about the role of such chaotic models in
ecology (Bascompte & Sole! , 1995; Cohen, 1995;
Blarer & Doebeli, 1996; Kaitala & Ranta, 1996;
Sugihara, 1996; White, Begon & Bowers, 1996a;
White, Bowers & Begon, 1996b).
With the presence of low frequencies in population
time series now rmly established experimentally,
other questions arise: where do they come from? Are
they induced by external inuences from the
environment, which typically have signals rich in
low frequencies (Steele, 1985; see also Grieger, 1992,
and references therein)? Or are they produced by
the intrinsic dynamics of the population, i.e. by the
interactions between individuals ? Also, in what
proportion do low frequencies arise? Are they
strongly dominant, like in the 1\f # spectrum of the
Brownian signal ? Or is it less predominant like in the
1\f signal, sometimes called pink noise ? (see
Halley, 1996, for a discussion).
Interesting results which address these issues have
been published by Miramontes & Rohani (1998).
Their approach consists of analysing time series from
laboratory insect populations and extracting their
15000
10000
5000
0
0 100 200 300
I
n
s
e
c
t
p
o
p
u
l
a
t
i
o
n
Time t (days)
Fig. 15. Time series from a laboratory insect population of
Lucilia cuprina. Notice the change in patterns after
approximately 200 days of observation. Reproduced from
Miramontes & Rohani (1998).
low-frequency content. The insect population under
study is that of Lucilia cuprina, or Australian sheep
bowy, which Nicholson (1957) kept under identical
conditions with a constant food supply for a duration
of approximately 300400 days. Population density
was evaluated roughly daily, providing a data set of
360 points (see Fig. 15).
This data set has attracted an enormous amount
of attention in the literature. This is rst due to the
irregularities it exhibits, even under constant en-
vironmental conditions. Another interesting feature
of the data is the change in the behaviour of the time
series after approximately 200 days. This was
investigated by Nicholson (1957) who found that,
unlike wild or laboratory stock ies, female ies after
termination of the experiment could produce eggs
when given very small quantities of protein food (in
fact small enough that the original strains of ies
could not produce eggs with it). Mutant ies might
therefore have appeared in the colony and, because
they were better adapted to living in such close
quarters and with a limited food supply, took over
the whole population. The eect of this change in
the insect colony can be seen in the behaviour of the
population density which uctuates more and
becomes more ragged. For a theoretical investigation
of this phenomenon using non-linear models, see
Stokes et al. (1988).
The population time series of Nicholsons (1957)
experiment obviously contains low frequencies and
long trends. To better quantify this content, Mira-
montes & Rohani (1998) applied the three methods
outlined in Section II.3 to the population density of
184 T. Gisiger
Lucilia cuprina, and also to that of the wasp parasitoid
Heterospilus prosopidis and its host the bean weevil
Callosobruchus chinensis, cultured by Utida (1957).
The results from their analyses are consistent with a
1\f structure of the noise, rather than an 1\f #
structure. They also nd a power-law distribution
for the absolute population changes D(s) `s
with
between 2.8 and 1.7, and D() `
with between
0.95 and 1.23 for the distribution of the duration of
these uctuations. These studies show that the
intrinsic dynamics of an ecosystem, even one com-
prising a single species and without any external
perturbations, is able to generate long trends in
population density. It shows also that a 1\f power
spectrum seems to be favoured over redder signals.
This frequency dependence, and the existence of
power-law distributions of event size and duration in
the system, seems to hint toward a critical state,
instead of a chaotic one.
(b) Ecosystems as critical systems
We end this Section on ecology by presenting further
evidence suggesting that some ecological systems
seem to operate near a critical state. We start with
the investigation of rain forest dynamics by R. V.
Sole! and S. C. Manrubia (Sole! & Manrubia, 1995a,
b; Manrubia & Sole! , 1997), and nish with the work
of Keitt & Marquet (1996) on extinctions of birds
introduced by man in the Hawaiian islands.
The two main forces which appear to shape the
tree distribution in rain forests, which will be of
interest to us here, are treefall and tree regeneration.
There is also competition in the vegetation in order
to get as much sunlight as possible, inclining trees to
grow to large heights. From time to time, old trees
fall down, tearing a hole in the surrounding
vegetation, or canopy. Because of the intricate nature
of the forest, where trees are often linked to others by
elastic lianas, when a large tree falls, it often brings
others down with it. In fact, it has been observed
that treefalls bring the local area to the starting point
of vegetation growth. The gap in the vegetation is
then lled as new trees develop again. This constant
process of renewal assures a high level of diversity in
the ecosystem.
Gaps in the vegetation, which are easy to pinpoint
and persist for quite some time, can be gathered by
surveys and displayed on maps. Sole! & Manrubia
(1995a) chose the case of the rain forest on Barro
Colorado Island, which is an isolated forest in
Panama. They present a 50 ha plot showing 2582
low canopy points where the height of trees was less
10
3
10
2
10
1
10
0
10
0
10
1
Gap size s (m)
G
a
p
d
i
s
t
r
i
b
u
t
i
o
n
D
(
s
)
B
A
Fig. 16. (A) Plot of canopy gaps obtained from numerical
simulations of the rain forest model of Sole! & Manrubia
(1995a). Each dot represents a gap (zero tree height) in
the canopy. (B) Frequency distribution of gaps from the
Barro Colorado Island rain forest for the years 1982 and
1983. The dashed line, representing a power-law dis-
tribution D(s) `s
".(%, ts the data quite well. Re-
produced from Sole! & Manrubia (1995a).
than 10 m, in the years 1982 and 1983 (see Fig. 16A
for a similar plot obtained by numerical simulation
of the model described below). The map shows holes
of various sizes in the vegetation, scattered across the
plot, in a possibly fractal manner. To verify this, Sole!
and Manrubia (1995a) rst computed the frequency
of occurrence D(s) of canopy gap size s (see Fig 16B).
The distribution ts a power law with exponent
k1.74 quite well, showing that there does not
appear to be any typical size for gaps. Another
indication of the fractal nature of the gap distribution
can be gathered by computing the fractal dimensions
of this set. Using methods such as the basic box
counting method presented in Section II.2, the
authors found the non-integer value 1.86. Sole!
and Manrubia (1995a) show in fact that, typical of
real fractals, the rain forest gaps set possesses a whole
spectrum of fractal dimensions which shows corre-
185 Scale invariance in biology
lations of the gaps on all scales and ranges : it
therefore seems to be a large, living fractal structure.
The presence of fractals and power-law distri-
butions are strongly suggestive that the rain forest
has evolved to a critical state where uctuations of
all sizes are present. To verify this hypothesis, Sole!
and Manrubia (1995a, b) propose a simple, critical
mathematical model which reproduces some aspects
of the dynamics.
The forest is represented using a generalization of
the model for plant growth of Section III.3 which
implements tree birth and death by four rules. Trees
start to grow in vacant areas using the stochastic
mechanism of Section III.3 (rule 1). Rule 2
implements tree growth at locations where the
surrounding vegetation is not taller, so as to
reproduce the eect of light screening by larger trees
on smaller ones. Spontaneous tree death can take
place for two reasons : because of age or because of
disease. Rule 3 implements these mechanisms by the
introduction of a maximal value for tree age (after
which the tree dies) and a random elimination (with
a small probability) of trees of all ages. Finally, rule
4 reproduces the eect of treefall on the surrounding
vegetation: when a tree dies and falls, it also brings
down some of its neighbours. This rule takes into
account as well the fact that older trees are higher
and will therefore damage a larger area of canopy
than smaller ones. We can see that rules 2 and 4 are
especially important to the dynamics of the system
because they introduce spatial and temporal corre-
lations in the system.
During the simulations, the system evolves as the
four rules are applied successively at each time step.
In doing so, the dynamics develops correlations
between dierent points of the forest, and nally on
all length and time scales. The latter is observable by
studying the time series given by the variation of the
total biomass of the ecosystem. The power spectrum
of the signal reveals a 1\f
dependence on frequency
with 0.87 ~~1.02. Spatial correlations can be
studied by computing the fractal dimensions of the
gap distribution in the system. Sole! & Manrubia
(1995a, b) and Manrubia & Sole! (1997) found
results quite similar, both qualitatively and quan-
titatively, to those obtained from the real data set,
giving strong arguments in favour of the rain forest
operating near or at a critical state. We refer the
interested reader to their articles for further details
and applications of their models.
Keitt & Marquet (1996) approach the study of
the possible critical nature of ecosystems from a
dierent angle, focusing instead on their dynamics as
they are gradually lled by species. For this task,
they chose the geographical area formed by six
Hawaiian islands : Oahu, Kauai, Maui, Hawaii,
Molokai and Lanai. These islands were originally
populated by native Hawaiian birds, until they were
driven extinct by Polynesian settlers and afterwards
by immigrants from North America. Since then,
other bird species have been articially introduced.
Records show that 69 such species were introduced
between 1850 and 1984. They also contain data
regarding the number of extinctions that occured
during this period (35 species went extinct during
these 70 years). Keitt & Marquet (1996) analysed
these records and found several indications that the
system might be operating in a critical state.
Extinction events seem not to have occured until
more than eight species were introduced into the
ecosystem. This transition does not happen in a
continuous manner, but more in an abrupt fashion,
reminiscent of a phase transition. Keitt & Marquet
(1996) interpret this as the system going from a non-
critical state to a critical one. Also, the number of
extinction events seem to follow a power-law
distribution, with an exponent around k0n91. This
means that small extinction events are a lot more
common than larger ones. We will see below(Section
IV.2) that similar power laws are suggested by the
analysis of fossil records. Lastly, the distribution of
the lifetimes of species, which range from a few years
to more than 60 years, also follows a power law with
exponent k1.16. These ndings might therefore
illustrate how an ecosystem self-organizes into a
critical state as the web of interactions between
species and individuals develops. However, more
data on this or similar ecosystems might prove
valuable to support this claim.
(2) Evolution
In this Section, I will present the result of some recent
work done on evolution using mathematical models.
There has been a great deal of activity in this area in
the last 10 years or so, mainly because of exciting and
unexpected patterns emerging from fossil records for
families and genera. These patterns, which are best
described by power laws and self-similarity, give
biologists solid data to build models and better
understand the mechanisms of evolution and selec-
tion at the scale of species.
I will start by presenting some of the power laws
extracted from the fossil records (Section IV.2.a).
This will be followed by a few remarks about models
186 T. Gisiger
70
60
50
40
30
20
10
0
70
60
50
40
30
20
10
0
500 400 300 200 100 0
Geological time (Myr)
P
e
r
c
e
n
t
e
x
t
i
n
c
t
i
o
n
P
e
r
c
e
n
t
o
r
i
g
i
n
a
t
i
o
n
A
B
Fig. 17. Time series for the percentage of origination (A)
and extinction (B) compiled from the 1992 edition of A
Compendium of Fossil Marine Families (see text). Reproduced
from Sepkoski (1993). Myr, million years before recent.
in evolution, and the concepts and notions they use
(Section IV.2.b). I will then present models proposed
to reproduce these measurements (Sections IV.2.ce).
(a) Self-similarity and power laws in fossil data
Traditionally, evolution is viewed as a continuous,
steady process where mutations constantly introduce
new characteristics into populations, and extinctions
weed out the less-t individuals. Via reproduction,
the advantageous mutations then spread to a large
part of the population, creating a new species or
replacing the former species altogether, therefore
inducing evolution of species. This mechanism,
named after Darwin, is in fact a microscopic rule of
evolution, or microevolution (i.e. it plays at the level
of species or individuals), and it governs the whole
ecosystem from this small scale. According to this
mechanism one expects extinction records to show
only a background activity of evolution and ex-
tinction, where a low number of species constantly
emerge and others die out.
However, fossil records show that history has
witnessed periods where a large percentage of species
and families became extinct, the so-called mass
extinction events . The best documented case is the
annihilation of dinosaurs at the end of the Cretaceous
period, even though at least four such events are
known (see Raup & Sepkoski, 1982 for a detailed list
of these events). These events clearly cannot be
accounted for in terms of continuous, background
extinctions as they represent discontinuous com-
ponents to the fossil data.
To explain these occurrences, events external to
the ecosystems were introduced. It is known that the
earths ecosystems have been subject to strong
pertubations such as variations in sea level, world-
wide climatic changes, volcanic eruptions and
meteorites, to name a few. Although the eect of
these events on animal populations is not very well
understood, it is quite possible that they could have
aected them enough to wipe out entire species and
genera (Homan, 1989; Raup, 1989). The record of
the extinction rate as a function of time should then
consist of several sharp spikes, each representing a
mass extinction event, dominating a constant back-
ground of low extinctions as diversication is kept in
check by natural selection (see g. 1 of Raup &
Sepkoski, 1982 for data with such a structure).
People interested in the interplay of evolution and
extinction have therefore traditionally pruned the
data, subtracting from it mass extinction events and
any other contribution believed to have been caused
by non-biological factors. However, with the ac-
cumulation of fossil data and their careful systematic
analysis, a somewhat dierent picture of evolution
and extinction has developed recently.
The rst such study was carried out by Sepkoski
(1982) in A Compendium of Fossil Marine Families
which contained data from approximately 3500
families spanning 600 million years before recent
(Myr). This was recently updated to more than 4500
families over the same time period (Sepkoski, 1993).
This record enables one to see the variation in the
number of families as a function of time, as well as
the percentage of origination and extinction (see
Fig. 17). As can be seen in Fig. 17B, the extinctions
do not clearly separate into large ones (mass ex-
tinctions) and small ones (background extinctions).
In fact, extinctions of many sizes are present : a
few large ones, several medium-sized ones and lots
of small ones. This characteristic of the distribution
seems to be robust as was shown by Sepkoski (1982),
being already present in the 1982 Compendium. An-
other striking fact is that the origination curve
(Fig. 17A) is just as irregular as the extinction
curve.
187 Scale invariance in biology
300
0
60
40
20
0
500 400 300 200 100 0
Geological time (Myr)
P
e
r
c
e
n
t
f
a
m
i
l
y
e
x
t
i
n
c
t
i
o
n
T
o
t
a
l
f
a
m
i
l
y
e
x
t
i
n
c
t
i
o
n
A
B
200
100
600
Fig. 18. Time series of family extinction for both marine
and continental organisms, as compiled from Fossil Record
2 (Benton, 1993). (A) Total extinction of families. (B) Per
cent extinction of families. Each graph contains a minimal
and maximal curve, meant to take into account un-
certainty in a variety of taxonomic and stratigraphic
factors. Reproduced from Benton (1995). Myr, million
years before recent.
A similar result was obtained by Benton (1995)
using the Fossil Record 2 (Benton, 1993), which
contains 7186 families or family-equivalent taxa
from both continental and marine organisms. Fig.
18A shows the total number of family extinctions
and Fig. 18B the percentage of family extinctions as
a function of time. These curves too show extinctions
of many sizes. There are also similarities between the
general shape of the extinction curves from the fossil
compilations of Sepkoski (1982) and Benton (1993),
even though the curves correspond to organisms
which lived in dierent geographical areas. This last
fact had been noticed by Raup & Boyajian (1988),
using 20000 specimens of the 28000 from the Fossil
Compendium of Sepkoski (1982). They examined the
similarities between the extinction curves belonging
to dierent families or species and found that they
were quite similar, even if the species or families
concerned lived very far fromeach other. To describe
this situation, Raup & Boyajian (1988) coined the
60
0
10
2
Per cent extinction
N
u
m
b
e
r
o
f
e
x
t
i
n
c
t
i
o
n
e
v
e
n
t
s
A
B
50
40
30
20
10
0 20 40 60 80 100
10
1
10
0
10
0
10
1
10
2
Fig. 19. (A) Distribution of extinction events as a function
of their size, for 2316 marine animal families of the
Phanerozoic. Reproduced from Sneppen et al. (1995). (B)
Same distribution plotted on a loglog scale. The straight
line has slope 2.7 and has been tted to the data while
neglecting the rst point which corresponds to events
smaller than 10%.
phrase that the taxa seemed to march to the same
drummer. They concluded that this cannot ob-
viously result from purely taxonomic selectivity, and
that external, large-scale, non-biological phenomena
were responsible for most of these extinctions.
However, a closer inspection of the extinction
curves shows even more striking results. Raup (1986)
sorted the extinction events according to their size,
and computed the frequency of each size of events
(see Fig. 19A). This distribution is smooth, instead of
consisting of just two spikes corresponding to small
extinction events (background extinction) and very
large events (mass extinctions). In fact, it seems to
follow a power law as can be seen in Fig. 19B. The
frequency of the smallest extinction events (smaller
than 10%) is rather far from the distribution.
However, this can be explained by the fact that small
events are more sensitive to counting errors, as they
can be masked by background noise. The exponent
of the distribution can be evaluated to be between
k2.5 and k3.0. If indeed, the distribution obeys a
power law, then extinction events of all sizes could
188 T. Gisiger
25
0
25
Time (Myr)
N
u
m
b
e
r
o
f
f
a
m
i
l
i
e
s
A
B
20
15
10
5
300
20
15
10
5
50 100 150 200
200 100
Fig. 20. Number of families of Ammonoidea as a function
of time over a period of 320 million years, with a time
denition of 8 million years (A), and 2 million years (B).
The bottom graph therefore is an expansion of the framed
region of the top graph, but at a higher time resolution.
Similar features appear as the time scale is reduced,
implying self-similarity in the record. Reproduced from
Sole! et al. (1997). Myr, million years before recent.
occur: from one family or species, to extinction of all
the organisms in the ecosystem. As in the case of
avalanches in the sand pile, there would be no
typical size for extinction events. Mass extinction
events would be contained in the skew end of the
distribution, and the separation of extinction events
into mass extinctions and background extinction
might therefore be articial and debatable.
If extinction event statistics follow a power-law
distribution, then the time series of the number of
families present as a function of time should also
have some interesting properties. This is indeed the
case, as shown by Sole! et al. (1997). They rst
pointed out that there is self-similarity in the fossil
record of the families of Ammonoidea (House, 1989)
(see Fig. 20). Taking part of the time series, and
expanding it by improving the time denition shows
a structure similar to the original record. The record
therefore seems to be self-similar, or fractal. Sole! et al.
10
1
Frequency f
P
o
w
e
r
s
p
e
c
t
r
u
m
P
(
f
)
A
B
10
2
Percentage of family extinctions
Total number of family extinctions
10
1
10
2
10
2
10
1
Fig. 21. Power spectra P( f ) of the time series of Fig. 18 for
the total number of family extinctions (A) and of the
percentage of family extinctions (B). The dashed lines
correspond to a 1\f
, where is some
positive parameter smaller than 1 which denes the
frequency of mutations. When a species does mutate,
it is then assigned a new tness. q
i
is a non-linear
function of B
i
which ensures that only species with
low tness will mutate since those with high B
i
have
considerably fewer options to raise their tness
further. It is important to note at this point that, in
the BakSneppen model, extinction and evolution
are two facets of a single mechanism: the species with
lowest tness goes extinct and is replaced by a new
one, which evolved from the former by mutation
(therefore inducing evolution). Evolution therefore
takes place in the ecosystem only because an
extinction event has occured. This is a very simplied
view of the phenomenon which has been modied in
subsequent models.
Fig. 24A shows the state of the ecosystem at some
initial time where we have assigned to each species a
random value for its tness. Fig. 24B shows the result
of the application of a purely Darwinian rule of
natural selection, i.e. survival of the ttest . The
system converges to a state where all the tnesses are
close to 1. We note that the convergence to this state
becomes increasingly slower as the minimum tness
of the species present in the ecosystem is progressively
raised by selection: at the beginning of the simu-
lation, the system goes through many mutation
events, but they become less and less frequent with
time. This does not reproduce the patterns we saw in
Figs 17 and 18, with extinction events of all sizes over
the whole time record. The dynamics is clearly
incomplete. The interactions, especially dependence,
between species which were present in the NKC
model for instance, are missing.
To remedy this problem, Bak & Sneppen (1993)
implement the interactions between species of the
model in the following way: whenever a species
mutates, so will its immediate neighbours. The
addition of this simple rule gives a dramatically
dierent dynamics to the system. Instead of nding
itself again in a situation where B is close to 1 for
every species, the ecosystem settles in a dierent state
where almost all the tnesses have grouped them-
selves in a band higher than some threshold value B
c
0.6670 (see Fig. 24C). This state is stable, as it
194 T. Gisiger
1
0.5
0
1
0.5
0
1
0.5
0
0 100 200
Species
F
i
t
n
e
s
s
A
B
C
Fig. 24. Distribution of the values of the tness B for an
ecosystem of Nl200 species. (A) Initial state where all
tnesses are chosen randomly. (B) State of the ecosystem
after following the purely Darwinian principle of re-
moving the least t species and replacing it by a mutant
species. (C) The same ecosystem after the same number of
time steps but using also the updating rule of BakSnep-
pen which takes into account interactions between species.
persists however long we let the simulation run for.
The convergence to this state gives the system the
following interesting dynamics.
As long as all species have tnesses above B
c
, the
system is in a phase of relative tranquility where
mutations seldom happen. Typically, one should
wait a period of approximately 1\q
c
le
B
c
/
10#*
time steps to see a single mutation occur. Here, the
organisms coexist peacefully and there is little change
taking place in the ecosystem.
However, when a mutation does occur in which
one of the species inherits a tness lower than B
c
, the
system enters a phase of frantic activity as the
probability of a mutation taking place is now much
higher than q
c
. This state is able to sustain itself as
15000
10000
5000
0
75
50
25
0
0 250 500 750 1000
Time (arbitrary units) A
c
c
u
m
u
l
a
t
e
d
c
h
a
n
g
e
(
a
r
b
i
t
r
a
r
y
u
n
i
t
s
)
S
i
z
e
o
f
e
x
t
i
n
c
t
i
o
n
/
m
u
t
a
t
i
o
n
e
v
e
n
t
s
0 20 40 60 80 100
A
B
Fig. 25. (A) Time series of the size of extinction\mutation
events for an ecosystem with 20 species (in most events, a
single species can mutate several times). (B) Time
evolution of the accumulated change of a given species of
the ecosystem. Every time the species mutates, the
accumulated change increases by one unit. The time is
measured in mutation events.
species get knocked out of the high end of the tness
region when one of their neighbours mutates. The
system will eventually settle back momentarily to its
quiet state when all species again have BB
c
. The
series of extinctions\mutations which took place
forms an event of measurable size and duration. Fig.
25A shows a series of such events as a function of
time. One notices that there are events of many sizes
similarly to the records from Fig. 17. Bak & Sneppen
(1993) have shown that the distribution of event size
and duration follows a power law, as does the
distribution of interaction ranges between species of
the ecosystem. This demonstrates that the system has
indeed reached a critical state where species interlock
spontaneously (i.e. without any parameter ne
tuning) into a chain of codependence. What happens
is that in the quiet state, stress has been building up
in the ecosystem like in the sandpile when it is very
steep. When a species mutates to a tness lower than
B
c
, it is like adding the grain of sand which triggers
an avalanche. In this case, it is a co-evolutionary
195 Scale invariance in biology
avalanche where mutation of species induces further
mutations of their neighbours.
Fig. 25B shows the accumulated change of a
given species in the ecosystem as a function of time.
Notice the periods of change (vertical lines) sepa-
rated by periods of stasis (horizontal line). This
compares well with Fig. 22 illustrating the notion of
punctuated equilibrium.
For further details on the model, its dynamics and
how it compares with fossil data, the reader is
referred to the literature (Sneppen, 1995; Sneppen et
al., 1995).
Another interesting model which describes ecosys-
tems as critical system was introduced by Sole! (1996)
and Sole! & Manrubia (1996). Here, evolving
interactions between species is the key ingredient
which drives the system towards a critical state. At
each time step, the total stress imposed by the rest of
the ecosystem on each species is computed. Species
which are subject to too much pressure go extinct,
and are replaced by mutants of the remaining
species. There are also spontaneous mutations
occurring, but at a lower level. This will drive the
system towards a state where removing just one
species will perturb the existence of many others, and
extinction events of all sizes will occur. Sole! ,
Bascompte and Manrubia (1996) and Sole! and
Manrubia (1997) showed that this system has
become critical with an extinction event distribution,
among other things, which follows a power law. For
further details see Sole! et al. (1996), Sole! &Manrubia
(1997) and references therein.
(e) Non-critical model of mass extinction
So far in our developments, we have presented work
which interprets the presence of power laws in fossil
data as evidence that ecosystems operate at, or near,
a critical state. However, power laws do not always
imply criticality as Newman (1996, 1997a, b)
showed. Indeed, he proposes a model for evolution
where no interactions are explicitly present between
species (although they may be implicit) which
accounts well for the data and seems to indicate that
codependence between species might, although
important, not be an essential ingredient in eco-
system dynamics. As we will see, his model therefore
does not interpret mass extinction events as co-
evolutionary avalanches, but rather as the result of
strong external stress on a temporarily weakened
ecosystem.
Part of the model of Newman (1997a, b) is the
Darwinian mechanism of elimination of unt species
by selection, and which is implemented as follows. As
in the BakSneppen model, each species possesses a
tness B. Stress, symbolized by a number between 0
and 1, is drawn using a random number generator
and applied to the system, eliminating all species
with tness smaller than this number. This stress can
be of physical origin (geographic location, climate,
volcanic eruptions, earthquakes, meteorites, etc.)
but also it can be caused by other species (predators,
competitors, parasites, etc.). Selection is then fol-
lowed by diversication as free ecological niches (in
this case, the spaces left vacant by the species which
just disappeared) are lled by new species. However,
these dynamics of a purely Darwinian type lead to
the same situation as the one illustrated by Fig. 24B
where the ecosystem becomes increasingly stable by
lling itself with highly t species. The interesting
variation which Newman (1997a, b) introduces to
complement his model is, instead of codependence,
the spontaneous ability of species to mutate and
evolve at all times, even in times of little external
stress. A new rule is then added to the model which
allows at each time step a small proportion of the
species of the ecosystem to mutate spontaneously.
This changes radically the dynamics of the system
which favours now a rich and complex activity.
Simulations show that the distribution of the size of
extinction events, as well as of species lifetimes obey
power laws. By making supplementary assumptions
about the model, Newman (1997a) was also able to
estimate the distribution of species in each genus,
which is also of this type and in the spirit of the
ndings of Burlando (1990, 1993) (see Section
IV.2.a).
That this model is able to produce power laws
without any parameter ne tuning or resorting to a
critical state is quite impressive. It is important to
note that the ecosystem of Newmans model (1997a,
b) is not critical : the species do not spontaneously set
themselves in highly unstable states to then become
extinct like falling dominoes. These events, called
coevolutionary avalanches, are a trademark of
critical models of evolution and a prediction which
has to be tested against fossil data. In Newmans
model (1997a, b), on the other hand, there are no
such avalanches as there are no direct interactions
between species. Interactions are implicitly included
in the stresses applied to the system. However, the
particular dynamics of the system allows it to make
original predictions of its own.
The model rst predicts that the longer the
waiting time for a large stress, the greater the next
extinction event will be. Indeed, periods of low
196 T. Gisiger
100
75
50
25
0
0 100 200 300
Time (arbitrary units)
N
u
m
b
e
r
o
f
e
x
t
i
n
c
t
i
o
n
s
Fig. 26. Time series of extinction events for the model of
Newman (1997a, b) for an ecosystem of Nl100 species
illustrating the notion of aftershocks following a large
extinction event.
stresses have a tendency to increase the number of
species with low to medium tness. These will then
be wiped out at the next large stress. The model
therefore tells us that in order to adapt to their
surroundings, species can sometimes render them-
selves vulnerable to other, less frequent stresses. This
could be tested using fossil data if information about
the magnitude of external perturbations can be
obtained independently.
Another prediction of the model is the so-called
aftershock, illustrated by Fig. 26. After a large
extinction event, the ecosystem is almost entirely
repopulated by new species. Statistically, about half
of them will have tness lower than 0.5. If another
large stress is applied to the system, most of these new
species will be immediately wiped out as they do not
have time to evolve suciently to withstand this new
perturbation. The second event will therefore be of
large size because the preceding extinction event left
the ecosystem vulnerable. Newman (1997a) gave the
name aftershock to this phenomenon. This pre-
diction could also in principle be tested using fossil
data. We refer the reader to the literature for further
details (Newman, 1997a, b).
V. DYNAMICS OF EPIDEMICS
We now turn to the work of C. J. Rhodes, R. M.
Anderson and colleagues on type III epidemics.
Section V.1 presents evidence for scale invariance in
the dynamics of this type of epidemic. A self-
organized critical model, formerly introduced to
account for the spreading of forest res, and which
reproduces well experimental data is then discussed
in Section V.2.
(1) Power laws in type III epidemics
Let us consider the records of measles epidemics in
the Faroe Islands in the north-eastern Atlantic, and
more specically the time series of the number of
infected people in the community month by month,
also called the monthly number of case returns. This
example is interesting in several respects. The
population under investigation is approximately
stable, at between 25000 and 30000 individuals. It is
also somewhat isolated, the main contacts with other
communities taking place during whaling and
commercial trade. It is believed that this is the main
route by which the measles virus enters the popu-
lation. The virus therefore infects a, at least partially,
non-immune population and epidemic events of
various sizes take place. The population is, however,
small enough for the epidemics to die out before the
next one starts. Because of the easily noticeable
symptoms of measles, the record is also believed to be
highly accurate. Fig. 27A shows the time series of
measles case returns for 58 years, between 1912 and
1972, before vaccination was introduced. The record
shows 43 epidemics of various sizes and duration
(most of which are very small and do not appear on
the gure).
We dene an epidemic as an event with non-zero
case returns bound by periods of zero measles
activity. Its duration is the total number of
consecutive months during which case returns are
non-zero, and the size s of the event is the total
number of persons infected during that period. With
these denitions, the records show epidemic events
ranging from one individual to more than 1500 (and
engulng close to the whole islands in one instance),
and with duration between one month and more
than a year.
Rhodes & Anderson (1996a, b) and Rhodes,
Jensen & Anderson (1997) computed the size and
duration distribution of the events and obtained the
results shown in Fig. 27B, C. The cumulative size
distribution D(size s) clearly follows a power law
over three orders of magnitude, as does the duration
distribution of events over one order of magnitude:
D(size s) `s
a
, (17)
D(duration) `
b
, (18)
where a 0.28 and b 0.8. [The cumulative distri-
butions D(size s) `
a
and D(duration) `
b
197 Scale invariance in biology
1000
500
0
1900 1920 1940 1960 1980
Time (years)
10
2
10
1
10
0
10
0
10
1
10
2
10
3
10
2
10
1
10
0
10
0
10
1
Epidemic size s
Epidemic duration s
D
(
d
u
r
a
t
i
o
n
>
s
)
D
(
s
i
z
e
>
s
)
N
u
m
b
e
r
o
f
i
n
f
e
c
t
i
v
e
s
A
B
C
Fig. 27. (A) Monthly measles-case returns for the Faroe
islands (population of approximately 25000 habitants),
between 1912 and 1972. (B) Cumulative distribution
D(size s) of epidemics according to their size s compiled
from A. The straight line represents a power law with
exponent a 0.28. (C) Cumulative distribution D(dura-
tion) of epidemic event duration . The straight line
has a slope of b 0.8. Reproduced from Rhodes &
Anderson (1996b).
are related to the non-cumulative distributions D(s)
`s
and D() `
)
P
o
w
e
r
s
p
e
c
t
r
u
m
P
(
f
)
1/f
A
B
Fig. 30. (A) Power spectrum P( f ) as a function of
frequency f of the error in the reproduction of a spatial
interval. Reproduced from Gilden et al. (1995). (B)
Distribution D() of time delays , or periods of
inactivity, for a typical neuron from the visual cortex of
the macaque macaca mulatta. The straight line has a slope
of k1.58. Reproduced from Papa & da Silva (1997).
proceeds as follows. Subjects are asked to reproduce
N times a given time interval, chosen between 0.3
and 10 s, by pushing a button on a keyboard. The
error e
i
, i l1, , N, is then recorded, interpreted as
a time series and its power spectrum computed. The
resulting power spectrumP( f ) (see Gilden et al., 1995
for details) behaves like 1\f
form,
with ranging from 0.7 to 0.9. This raises the
hypothesis that 1\f spectra might be common to all
conscious natural behaviour. Further investigations
are certainly indicated.
(3) Scale invariance in the activity of neural
networks
We end this Section on scale invariance in the
central nervous systemwith some interesting ndings
on the ring of neurons in the cortex. In a recent
article, Papa & da Silva (1997) present a plot of the
distribution of time intervals between successive
rings of cortex neurons (Fig. 30B). The data used
for this analysis come from a study of the visual
cortex of the macaque macaca mulatta by Gattass &
Desimone (1996). As can be seen from Fig. 30B, the
203 Scale invariance in biology
1000
10
2
10
1
10
0
10
0
10
2
Time interval (ms)
D
i
s
t
r
i
b
u
t
i
o
n
D
(
)
A
B
10
1
10
1
10
3
750 500 250 0
Time (ms)
Fig. 31. (A) Firing response of a neuron from the visual
cortex of the cat when exposed ve separate times to the
same stimulus. Each vertical bar represents a spike: their
exact timing changes widely from one series to the next.
Reproduced from Koch (1997). (B) Distribution of the
time elapsed between two consecutive spikes in the record
shown in A, as a function of its frequency. The straight
line has a slope of approximately k1.65. The distribution
was obtained by compiling the time delays in a
histogram with 30 channels, giving a time resolution of
approximately 8 ms.
distribution D() of time intervals clearly follows
a power law D() `T